GuShiio.com Wen| Narrow Broadcast, author| Li Wei
Will AI search attempts driven by big models shake the dominance of traditional search engines? This question has been raised since ChatGPT was released, but there has been no clear answer.
Although the new Bing, which integrates large-scale model capabilities, exceeded 100 million daily active users a few weeks after its release, and AI search products such as Perplexity and Mita are also sought after by capital, they have not brought more urgency to the search business of Google and Baidu.
Google opened the AI search project Search Labs in May 2023. Users can apply to join the waiting list and wait to check the test invitation. The AI Overviews search feature was launched a year later and was initially only available in the United States. The launch of this feature is also known as“One of the biggest changes in Google’s search engine in 25 years.”
Baidu has been making attempts on AI search outside of Baidu search. First, in October 2023, we upgraded our simple search to AI interactive search with large model reconstruction. Then, in September 2024, the Wenxinyiyan App was updated to Wen Xiaoyan and positioned it as a new search intelligent assistant.
The popularization of deep thinking capabilities promoted by DeepSeek open source has intensified the intensity of competition.To a greater extent, it has shaken the old power pattern in the search field.
On the one hand, the integration of in-depth thinking capabilities allows AI to decompose, extend, correct and explore problems, improve information acquisition and integration capabilities, greatly optimize the question-and-answer search experience, and help it achieve a greater degree. Popularization of users; on the other hand, pressure from DeepSeek has also stimulated the explosion of various reasoning models, and has allowed closed-source manufacturers such as OpenAI to begin to lower the threshold and payment requirements for search products.
The development of big model technology has accelerated the innovation of search experiences, which in turn has begun to change the way people obtain information and habits.This change has weakened the matching efficiency advantages that traditional search engines have accumulated over years.
Under the traditional search experience, people are accustomed to searching and matching hyperlinks to information through keywords, and then reading, filtering, and sorting out the required answers. Under the AI search experience, more and more users are beginning to choose to communicate with AI through natural language, filter information and organize answers with the help of AI. People’s ultimate expectation for search products has changed from a library that links everything to a knowledgeable elderly person who can give any answer.
It is not only traditional search engines that are qualified to become knowledgeable eldersThere are also large model companies that master core technologies, AI search companies with product innovation capabilities, and platform companies with high-quality content reserves. Moreover, the changes in search are still in the early stages of exploration, technological progress is still the dominant factor in this competition, and more start-up teams are trying to seize blockbuster opportunities with new technological breakthroughs.
Traditional search engines have already felt this change. In the major version update on February 24, Baidu App fully launched the “AI Search” portal, and replaced the Slogan with “Baidu, you will know.”This change is Baidu’s determination to demonstrate to the outside world its determination to maintain a stable search boundary, and to a certain extent, it has become an example of the search power landscape becoming increasingly loose.
Despite the large number of players, in the long run, AI search needs to find the best product experience based on technological progress and establish a business model that matches it. In this process, game participants will find their own positions based on different longboard advantages and complete the reshaping of the search power landscape in the AI era.
Deep thinking accelerates AI search to break the circle
In a recent interview with Tencent Technology, Zhu Xiaohu said that most questions were searched using ChatGPT or chat robots like DeepSeek. He believes that before, he needed to enter a very accurate prompt to get a reply from the AI, and the experience was not good; now, by entering a very simple question, he can get a long list of replies from the AI, and he can continue to ask,”It is enough to satisfy my needs for obtaining information.”
It is the popularization of in-depth thinking ability that has brought about the improvement of the search experience in Zhu Xiaohu’s words. The realization of this in-depth thinking ability depends on the advancement of reasoning models. The key is that through the combination of reinforcement learning technology and CoT (Chain of Thought) technology, the large model has an information processing process similar to the human thinking process.Information acquisition or search is the first common scenario that can leverage the advantages of this technology.
OpenAI launched its first reasoning model o1 in September 2024, and introduced the technical principles of o1 in the blog post “Learning to Reason with LLMs”: before responding to user needs, o1 will generate a long internal thinking chain. Through intensive learning, O1 will continue to conduct self-questioning and reflection, dismantle problem-solving steps and replacement methods, and improve and improve strategies.
However, OpenAI has chosen a closed-source strategy, and users need to pay a monthly subscription fee of US$200 to experience o1 ‘s reasoning capabilities, which greatly limits the popularity of the experience. DeepSeek opens the R1 model and exposes the complete thinking chain in the deep thinking mode, showing the charm of the reasoning model to users around the world. It completed the detonation of market volume and user scale in a short period of time, and opened the door to in-depth thinking. The shackles of popularization of ability.
This also allows a large number of users to experience the intergenerational changes of AI search like Zhu Xiaohu.With the support of in-depth thinking ability, fuzzy questions raised by users will be disassembled and improved by the large model into a complete problem solving chain, and more comprehensive and accurate answers will be given step by step. Although there are still some illusions, on the whole, it can still bring great surprises to users.
Driven by technological breakthroughs and traffic explosion, the ability to think in depth has also become an important means for pan-search products to gain growth. For a while, AI search products, AI assistants, content platforms, and traditional search engines all connected to DeepSeek. Among them, WeChat launched grayscale test AI search at this point in time. While amplifying the market volume of Tencent’s AI business, it also combined with Yuanbao App to truly achieve universal support for the in-depth thinking experience.
This has stimulated large models such as Gemini, OpenAI, Claude, Tongyi Thousand Questions, and Hunyuan to accelerate the introduction of new reasoning models with in-depth thinking capabilities.Their own reasoning models have also become one of the options for these manufacturers to establish search advantages。At the same time, the diversification of reasoning models can provide AI search with more options for model capabilities, which in turn may lead to a better AI search experience and bring greater variables to AI search.
The search power landscape is still dynamically changing
AI search has the potential to become the first “killer application” in the AI field, can change the way people obtain and use information. The expansion of expectations and ambitions will continue to wash away the power landscape in the search field. In this wave of AI search popularization led by in-depth thinking capabilities, the current position is the respective starting line for companies, which determines their lower limit in this game.
First, the challenged traditional search engines have already felt the crisis, but they still have a stronger search mentality and are embracing AI search more actively.With the support of sufficient excellent organizational capabilities, traditional search engines are still strong contenders for the Iron Throne of AI search in the future.
StatCounter data shows that Google’s search engine market share will remain below 90% in the fourth quarter of 2024. Since April 2015, Google has maintained a market share of more than 90-92%. In a February 2024 report, Gartner expected traditional search engine traffic to fall by 26% in 2026, and now it seems that this number is likely to become even larger.
Innovation has become an inevitable choice for traditional search engines to maintain absolute market dominance.At a recent earnings conference, Google CEO Sandar Pichai said: “2025 will be one of the years with the largest search innovation to date.” As a manifestation of innovation, Google Search has been exposed and is developing an AI model. This pattern is supported by a customized version of the Gemini 2.0 model, has advanced reasoning and deep thinking functions, and can support questioning.
Baidu Search, which also feels threatened, also announced that it will fully access the latest in-depth search functions of DeepSeek and Wenxin Model after WeChat launched AI search, and will be open to users for free use. On this basis, Baidu also plans to combine the accumulation of RAG technology with DeepSeek to establish differentiated advantages in eliminating illusions, providing expert search services and expanding user scenarios.
Second, compared with traditional search engines that came from the old era, AI assistants are indigenous people in the AI era, have a higher starting point, and are the strongest challenger to traditional search engines. Even the final outcome of traditional search engines may evolve into AI assistants.
If the AI assistant is assumed to be a human, the search corresponds to the human’s information acquisition ability. The previous information acquisition function of AI assistant products was considered trivial. Even ChatGPT and Kimi, which had relatively better experiences, had difficulty meeting complex information acquisition needs.
After the in-depth thinking ability was launched, AI assistants such as ChatGPT, Kimi, Yuanbao, and Qwen basically have the in-depth thinking + online search function, which can retrieve corresponding links from public information on the Internet and integrate the content into more detailed answers.
AI assistants are closer to people’s way of thinking, and also carry the pan-search needs of more and more people through dialogue questions and answers.
This also means thatAI search will become the basic ability for AI assistants to attract new and promote activities。Previously, AI assistants had experienced a round of streaming battles. Not only did their retention effect average, but they were also overtaken by DeepSeek, whose performance improved. This prompts AI assistants to return to the path of growth in capabilities. andAI search is currently the ability with the largest user coverage and the clearest improvement path.
We can see that OpenAI has opened the AI search tool SearchGPT to everyone through ChatGPT, and users who are not registered can also use this search function. ByteDance’s AI assistant Doubao has also begun to test the deep thinking model on a small scale. Some users can see the thought chain in Doubao’s answers. Ali has launched a new Qwen product, which has AI search capabilities supported by the qwen2.5 model.
Third, the ticket for platform-based products to participate in the AI search battle is the accumulation of users or content in the mobile Internet era. This advantage will also likely be continued in the AI era, allowing these platform-based products to have a deeper say in the field of AI search.
The platform-based products here refer to products that have occupied a place in the mobile Internet era and have a considerable amount of content accumulation and user accumulation, such as WeChat, Douyin, Meta, X, Xiaohongshu, Zhihu, etc. These products have developed varying degrees of search habits internally.
WeChat and Douyin have been regarded as challengers to traditional search before this, and AI search has brought new opportunities to them.Relying on access to DeepSeek, WeChat has completed a large-scale cultivation of users ‘search habits.
Similarly, Meta is also expanding the Internet search function of Meta AI, signing up for more authoritative content sources, and strengthening search capabilities. Meta Chief Financial Officer Susan Lee believes that Meta AI will handle an increasingly wide range of search queries in the future.
Xiaohongshu and Zhihu have previously developed their own AI search products. Benefiting from the influence of open source of the DeepSeek model, Xiaohongshu and Zhihu can narrow the gap in model capabilities with giants, exert their content potential to a greater extent, and amplify their potential in the field of AI search. I know that Hu Direct Answer has been connected to DeepSeek, and Xiaohongshu is also connecting it for Diandian.
The integration of in-depth thinking capabilities improves the efficiency of content circulation on the platform and further strengthens users ‘search habits within the platform.On this basis, platform-based products are also expected to open up the introduction of more external resources and capabilities to meet the broader search needs of users.
Fourth, startups focusing on AI search have different starting points and also present three different development directions. Different directions point to different positions in different search power landscapes.But all startups have not escaped the threat from the first three categories.
Some AI search products are still committed to continuing to strengthen their AI search capabilities and continue to participate in the competition of general search. Leading AI search products such as Mita and Perplexity entered the market earlier, bound with more resources, and will have a more determined attitude. While smaller entrepreneurial teams such as Felo and Hika are trying to seize the general AI search capabilities, they also have the possibility of turning to the other two directions based on the accumulation of their own product capabilities.
There are also some AI search products that choose to focus on vertical areas and integrate large model capabilities and vertical information into their own competitive barriers. Sequoia Capital mentioned Harvey for the legal field and OpenEvidence for the medical field in the report, and believes thatSearch will no longer be a single market, but a diverse ecosystem of professional tools.
The last part of AI search products choose to build themselves into a hybrid of AI search + knowledge base. For example, ima, a subsidiary of Tencent, has a knowledge base function for storing documents, as well as AI search and AI writing, and opens pages associated with search results like a browser. In addition, there is Glean, which provides knowledge management and AI search services within enterprises. Its ARR (annual recurring revenue) in 2024 has reached US$55 million.
There are still capital that will be optimistic about the opportunities of start-up teams in the field of general AI search. The latest news is that AI search product Genspark has just completed a US$100 million Series A financing, with a valuation of US$530 million and 2 million users.The business story of starting from the beginning and turning into a giant is more legendary, which is also a natural charm of startups.
Four abilities determine the upper limit of development
Technical equality will bring equal opportunity to a certain extent.Under this dynamic and changing landscape, who is more likely to win depends on who can build the most solid balance between the four capabilities of model, product, content and commercialization.
The continuous improvement of model capabilities is the most critical condition for achieving a better AI search experience.The emergence of large models has accelerated the advancement of search technology, shifting search from simple web page indexing to semantic understanding and knowledge integration. The development of inference models has enhanced the ability of large models to understand requirements and integrate content, and refreshed people’s understanding of the upper limit of AI search capabilities.
Like DeepSeek,If the capabilities of the large model are strong enough, it is likely to have crushing advantages over other products, but this is too difficult to achieve.。At present, AI search players with the ability to develop large models themselves are not only more actively promoting the research and development of model capabilities, but also beginning to establish higher competitive barriers through closer integration of models and needs.
Public information shows that behind the Deep Research feature launched by OpenAI is supported by a customized OpenAI o3 model, optimized for web browsing and data analysis. Anthropic’s optimization of the Claude 3.7 Sonnet model also focuses more on meeting the actual needs of enterprises for large language models.
Google will also try to leverage the multimodal capabilities of the model to continue its search advantages on the hardware side. Google has developed the Circle to Search function based on multimodal capabilities. Users can search for target information by drawing a circle on the screen. Currently, Google is preparing to expand this experience from within the Android ecosystem to Chrome and Google Apps for the iPhone.
If closed-source models achieve absolute lead, then these companies are very likely to reach the iron throne of AI search relying on model capabilities.
ButIf the capabilities of open source models can always be close to closed-source models, the weight of product capabilities will be greatly increased。
Perplexity, a leader in AI search product innovation, will gain a greater chance of winning. Perplexity’s search results presentation has become a reference template for many AI search products. Based on the introduction of the large-scale model capabilities of OpenAI, DeepSeek, and Anthropic, Perplexity’s advantage is its ability to cover more daily use scenarios through product innovation.
Minta’s recently launched research model is also a reflection of product advantages. This function adopts a collaborative architecture of small model + large model. With the support of DeepSeek, it completes the disassembly of requirements and the improvement of search paths, and then uses the self-developed model to complete the search and integration of information. Data released by Minita shows that in this method, hundreds of web pages can be searched and analyzed in 2-3 minutes.
Small model development can be regarded as an optimization of product experience to some extent.Felo and Hika are also pursuing the possibility of becoming a general AI search based on product experience optimization。Felo supports vertical search of the content of Xiaohongshu;Hika will focus on asking more questions and talking about the design of functions in the search results to get as close as possible to people’s habit of divergent thinking.
As mentioned earlier, platform-based companies with content accumulation already have a natural moat.Content-based platforms are first the resource providers of AI search, and then the players at the table.
A high-quality content pool can improve the effectiveness of AI search results. Google has previously been criticized for errors and ridiculous answers generated by AI search, and has had to improve the accuracy of answers by restricting user-generated content and satirizing the use of website content on the source of AI search content. At the same time, using media information as much as possible in answering is also a means to improve effectiveness.This also gives high-quality content sources stronger bargaining power.
At the same time, WeChat, which has public account content, Douyin, which has video content, and Xiaohongshu, which has experience notes, will all be considered important participants in the future search power landscape due to their strong content accumulation. However, among these platforms,Several companies with more complete and open ecosystems also have the possibility of incubating general AI search or AI assistants。This is also why WeChat’s AI search has such widespread attention and influence.
Innovation in commercialization capabilities is a problem that all AI search products have not yet solved. Finding solutions faster will increase the possibility of success.This may determine the survival and development capabilities of some AI searches, but it will not be a decisive factor affecting competition.
At present, the most common one is the membership payment model. Some AI search products will set free basic functions and paid advanced functions based on the strength of the model capabilities, the number of files uploaded, etc. At present, advanced features often give a certain number of free experiences to attract users to pay.
Perplexity is also experimenting with advertising models, looking for appropriate advertising spaces in the search results interface, such as extended suggested questions, positions on the edge of answers, etc. The revenue from these advertising spaces will be used to quell media dissatisfaction with its non-compliant citations. However, one of the competitive advantages that AI search has always emphasized is ad-free. Although Perplexity does not insert advertisements in its answers, it may still cause user dissatisfaction and become an attack point for competitors.
In the longer term, AI search may not be limited to providing users with information acquisition and integration capabilities, but will become a terminal for users to invoke Agent services, and a large number of Agnets will provide users with more professional vertical domain services.This is an expansion of the “big model + small model” framework and may also create a business model where platform and Agent revenue are shared.
If this model can be implemented, it may be able to paint a future power landscape for the search field: AI search will be fully integrated with AI assistants to form several main entrances. More AI search products we see now will either go to extinction, or they will transform into a part of the Agent and become a vertical capability provider at the main entrance.
